FAQ

Is chukei open source or fair-source?

chukei is fair-source under FSL-1.1-ALv2, not OSI open source at release time. The source is public and usable for permitted purposes, including internal production use, but the license blocks using chukei to offer a competing product or hosted service. Each release converts to Apache-2.0 two years after that release is made available.

Can chukei serve a wrong (stale) result?

The cache is built to make that an engineering impossibility rather than a hope: only deterministic reads are cacheable, writes invalidate affected tables, chunked results are never cached, and blame mode continuously re-executes a sample of cache hits against live Snowflake and counts mismatches. The SLO is zero; if it ever fires, disable the cache with CHUKEI_PLUGINS_CACHE_ENABLED=false and file a bug — queries keep flowing uncached.

Does chukei see my credentials?

They pass through to Snowflake verbatim. chukei never persists or logs them; session tokens live in memory only. This is enforced by a trace-level credential-leak audit in the test suite.

What happens if chukei crashes?

Clients pointed at it retry, then error — they do not silently fall back. A restart faster than the driver retry budget (~10s) is invisible, and existing sessions resume without re-login. Run with a restart policy (Docker --restart, k8s Deployment) and rehearse the rollback: point the driver host back at <account>.snowflakecomputing.com.

Why isn't my big query cached?

Large results are returned as chunk URLs that drivers download directly from cloud storage, bypassing any proxy — so chukei deliberately never caches chunked responses. They pass through unmodified.

Why did my query miss the cache?

Common reasons: non-deterministic functions (CURRENT_TIMESTAMP(), RANDOM()), a write to a referenced table since the last run, TTL expiry, or differing bind values. When in doubt chukei misses — that's the false-positive-intolerant design.

What's the performance overhead?

Deterministic Rust on the hot path; ~2ms p99 measured locally, with a +5ms budget enforced as an alert threshold (chukei_proxy_overhead_seconds).